A machine learning approach to predict CO2 diffusivity in liquid H2O over a wide pressure and temperature range

Abstract

This study presents a machine learning approach for predicting the diffusivity of CO₂ in liquid H₂O over a wide range of temperatures and pressures. A comprehensive experimental dataset is compiled, including over 300 data points from existing literature, as well as, 75 newly identified diffusivity measurements. These data span a broad spectrum of temperatures and pressures. Various machine learning models namely, Support Vector Machines (SVM), Random Forest (RF), k-Nearest Neighbors (kNN), and Autoencoders, are trained on this enhanced dataset and evaluated for their accuracy in diffusivity prediction. Results show that the Autoencoder model achieves superior performance, accurately predicting CO₂ diffusivity even in regions where experimental data is sparse. The model’s ability to generalize across a wide range of temperatures and pressures, demonstrates its potential for use in real-world applications, enabling fast, reliable predictions with minimized computational cost.